This repository contains the implementation of our CS577 project, where we explore improving the robustness of AI-generated text detection by using adversarial learning, inspired by the RADAR framework.

Original paper reference: RADAR: Robust AI-text detection via adversarial learning

Recent advancements in large language models (LLMs) have made it increasingly difficult to distinguish between machine-generated and human-written text. This project aims to:

• Investigate the robustness of AI-text detectors under adversarial paraphrasing.
• Develop a hybrid paraphrasing strategy combining backtranslation and neural paraphrasing.
• Train a distilBERT-based detector to classify texts as AI-generated or human-written.

• Hybrid paraphrasing pipeline: Combines multilingual backtranslation and lexical neural paraphrasing.
• Adversarial training loop: Reinforcement learning-based paraphraser competes against a binary classifier.
• Evaluation on real-world data: Includes a manually annotated dataset of LinkedIn posts.

• Python >= 3.9
• PyTorch >= 2.6.0 with CUDA support
• HuggingFace transformers & datasets
• NLTK
• Helsinki-NLP models

• Backtranslation: English → French → English using Helsinki-NLP.
• Neural paraphrasing: NLTK-based paraphraser to create more natural variations.

• Detector: distilbert-base-uncased binary classifier.
• Paraphraser: t5-small fine-tuned with PPO.

• Paraphraser generates samples to fool the detector.
• Detector learns from these new samples to improve classification.

• AUROC, Accuracy, F1, Precision, Recall.

git clone https://github.com/Subangkar/cs577-project.git
cd cs577-project

python3 -m venv venv
source venv/bin/activate  # On Windows use: venv\Scripts\activate

pip install -r requirements.txt

Run training:

python radar.py

Run evaluation:

python radar_evaluate.py